The AILIS Primer: A Proposed Layer Model for AI Systems¶

Version 0.1 - Draft for Discussion • August 2025

Proposal: Could an OSI-style mental model help us better understand and discuss AI systems? This document explores a possible 16+ layer framework for organizing our thinking about the AI stack.

0) Why explore a layered "AI stack" model?¶

Most diagrams compress AI into 4–7 broad boxes (infra → model → orchestration → app). This might hide what we're calling the under-served middle where real interoperability challenges exist: registry, routing, transport semantics, session/identity, and governance. The AILIS proposal explores whether more granular layers could help identify these gaps.

Clarity: Common vocabulary for planning and architectural reviews.
Interoperability: Clear layer boundaries invite specs and conformance tests.
Strategy: Spot duplication (crowded layers) and white space (under‑served layers).

1) The Layers (L0–L16)¶

We keep L0 for facility concerns and extend through L16 (Application). Cross‑cutting planes (Control, Management/Observability, Security) apply across layers.

L0 – Facilities & Power — Datacenters, power/cooling, physical security.
L1 – Compute Fabric — GPUs/TPUs/NPUs/CPUs, memory, interconnects.
L2 – System & Driver Runtime — CUDA/ROCm/Metal; device memory management.
L3 – ML Graph & Compilation — XLA/TVM/TensorRT‑LLM/ONNX Runtime.
L4 – Numeric & Quantization — FP16/FP8/INT4, sparsity, calibration.
L5 – Tokenization & Encoders — BPE tokenizers, CLIP, audio patchifiers.
L6 – Model Parameters & Architecture — Base/foundation weights, MoE, diffusion.
L7 – Inference Engine & Decoding — Serving runtimes, caching, speculative decoding.
L8 – Context Construction & Prompting — System prompts, templates, few‑shot.
L9 – Knowledge & Retrieval — Vector/graph indexes, rerankers, grounding, citations.
L10 – Tool & Function Invocation — Typed tool I/O (MCP), function calling, API bindings.
L11 – Addressing & Registry — Signed manifests, discovery, capability vectors, fingerprints.
L12 – Routing, Planning & Policy — Rule DSL + bandits; budgets, privacy, fallback/parallel.
L13 – Transport & Flow Semantics — Idempotent runs, streaming, CANCEL/RESUME, multiplex.
L14 – Session, Identity & Memory — Portable session envelope, capability tokens, budgets; memory tiers.
L15 – Governance, Safety & Schema — Redaction, validation/repair, schema change control, audit.
L16 – Application & Domain Logic — Product UX, workflows, agent frameworks.

Planes: Control (policy/config), Management/Observability (telemetry/evals), Security (mTLS, keys, PII).

2) Where the market is crowded vs. empty¶

Crowded: L1–L7 (compute, models, inference) and L9 (vector DBs).
Under‑served: L11–L15 (registry, routing, transport semantics, session, governance). This is where

multi‑provider, privacy‑aware, cost‑aware AI breaks down today.

3) What good looks like at the middle layers¶

L11 — Addressing & Registry¶

Objective: a portable, signed catalog for models, tools, indices, and ensembles.
Artifacts: manifests, capability vectors, immutable fingerprints (benchmarks + environment descriptors).
APIs: search/filter, publish/update, signature verification, evidence links.

L12 — Routing, Planning & Policy¶

Objective: pick the best plan (single model/tool or ensemble) under cost/latency/privacy budgets.
Mechanics: declarative rules + bandit learning; cascade/parallel/reducer; canary & rollback; evaluator feedback loop.

L13 — Transport & Flow Semantics¶

Objective: make AI calls resumable, cancellable, multiplexed, and stream‑aware across providers/hosts.
Verbs: OPEN, DATA, YIELD, NEGOTIATE*GRAMMAR, CANCEL, PAUSE, RESUME, CLOSE + idempotency keys and resume tokens.

L14 — Session, Identity & Memory¶

Objective: continuity and ownership of context across providers.
Artifacts: Portable Session Envelope (identity, budgets, privacy class, ensemble stack, memory refs) + capability tokens. Memory tiers: episodic (per thread), semantic (indices), scratch (structured workspaces).

L15 — Governance, Safety & Schema¶

Objective: reduce risk without blocking productivity.
Mechanics: input redaction; output validation/auto‑repair; schema diffing and deprecation windows; approvals & audit.

4) Ensembles: composition as a first‑class citizen¶

Real systems compose personas, prompts, templates, and skills. Ensembles are measured (capability vectors) and selected (routing policies) like any other capability. Evaluation must consider interaction effects—small changes can have large behavioral consequences.

5) Evidence over hype: fingerprints & capability vectors¶

Fingerprint = immutable record of (subject × platform × model‑version × env) performance: behavioral, performance,

robustness, safety, compatibility.

Capability vector = compact features for routing decisions (reasoning, factuality, structure adherence,

latency/cost bands).

Loop: route → score → action → outcome → update priors.

6) Security & privacy by design¶

Privacy classes & residency in manifests; privacy‑aware routing.
Content‑free telemetry option; on‑device redaction in clients; short‑lived capability tokens.
For premium IP: attested runs, watermark/provenance, rights metering.

7) How might AILIS be used in your organization?¶

Map products to layers (primary/secondary).
Identify gaps in L11–L15; pick standards to adopt/publish.
Instrument evaluation and publish fingerprints to your registry.
Write routing policy (start rules → evolve to bandits).
Govern schemas and introduce a firewall before tool/model calls.

8) FAQ¶

Is AILIS meant to replace existing frameworks? No. It's a proposed map and a set of ideas for discussion, not

a framework.

Why not fewer layers? The missing interoperability happens precisely in the “extra” layers (L11–L15).
Does it handle multi‑modal? Yes—modalities live in manifests/fingerprints; layers remain the same.

9) What’s next¶

Publish spec drafts for L11–L15 and the Portable Session Envelope.
Provide reference implementations (registry server, router DSL engine, transport shim, session libraries).
Invite vendors and users to contribute fingerprints and manifests.

*Feedback welcome._